Dragon Army Retrospective

Author’s note: I’m currently in the process of transferring a bunch of content from an archive over to Medium. The following is the retrospective from the six-month group house experiment I ran in Berkeley in fall 2017/spring 2018.


Final grade (counting up from zero, comparing what we tried and accomplished with the counterfactual Just Another Rationalist House): B-

Final grade (counting down from perfection, comparing us to the vision laid out in the Charter and the best realistically envisionable thing): D-

Final grade for Duncan: D
Final average grade for other Dragons: C+
Spread of grades for other Dragons: D-, D, C-, C, C, C, B-, B, B, B+

I’ve toyed with many different formats for an update/retrospective, and in the end I settled on advice to past me as the most relevant and useful lens to outsiders seeking to learn from Dragon Army (either to copy its successes or to avoid its failures). These are the tweaks and tinkers that I would send back in time to change the outcome of the experiment.

(I won’t be engaging with the question of whether-or-not-past-me-should-pull-the-trigger. The experiment was obviously worth running at the time, and I’m glad I ran it in retrospect, and I would run it again if teleported back in time. There was a particular subset of LessWrong and Tumblr that objected rather … stridently … to even considering something like Dragon Army —

(here I am specically not referring to those who pointed at valid failure modes and criticized the idea in constructive good faith, of whom there were many)

— those people were wrong then and they would be wrong now; I myself was wrong to engage with them as if their beliefs had cruxes that would respond to things like argument and evidence. Suffice it to say, the sky did not fall, and we never came remotely close to any of their collectively prophesied Dooms. I’ll be curious to see if any of them hold to the principle that one should be equally loud and enthusiastic in proclaiming “I was wrong,” but I’m not holding my breath. </snark>)


(Note: these will be all over the place; they aren’t in order, they aren’t all of the same order of magnitude; they aren’t even all of the same type; some will overlap.)

0. Know about, and tell the other Dragons about, stag hunts. Okay, I don’t know if I’m actually hitting the specific thing that is meant by stag hunts in the literature; there’s a post in the making on this that I hope will be useful to people and my version may be mangled from the original. But I think my version is super important, so in short:

All the players in the game have a choice of “stag” or “rabbit.”

All game brought down is shared evenly within the group.

Bringing down a stag is costly and effortful, and requires coordination. Let’s say you lose 5 utility to participate in a stag hunt, but if the group brings one down, it’s worth 100 utility, of which you’ll get 10.

Bringing down rabbits is low-cost and low-effort and can be done unilaterally. Let’s say you lose 1 utility to hunt rabbits, but you’ll typically bring in 3 utility by doing so.

If any player chooses rabbit, the stag escapes through the hole in the formation and is not caught. Thus, if 10 players all choose stag, they lose 50 utility and gain 100 utility, for net 50. If 9 players choose stag and one chooses rabbit, they lose 46 utility and gain 3. If all 10 players choose rabbit, they lose 10 utility and gain 30.

This creates a strong pressure toward having the Schelling choice be rabbit. It’s saner and safer if you have any doubt about the other hunters’ ability to stick to the plan, or the other hunters’ faith in the other hunters, or in the other hunters’ current resources and ability to even take a hit of 5 utility.

But stags, tho.

In retrospect, the whole frame of Dragon Army was me trying to build a coordination structure for choosing stag. Like in CFAR’s Hamming Circles, where after three days of suffering together through the rationality canon, we can ~100% reliably get groups of four people to all choose stag together, and fully show up to help one another with their Hamming problems. That’s not possible in regular day- to-day life, where casual defection and personal failure-to-model and all sorts of other Murphy-style things happen to punish the strategy of just-choose-stag. I wanted Dragon Army to be a place where people could trust one another enough to get the gains of going for a thing that can’t be got unless we’re all really truly in it together.

Having that frame at the outset would’ve had (I predict) something like a two-fold multiplier effect on our power as a group, and on my ability to make correct decisions as a leader.

1. Have “Dragon Army” be the house that starts after the first six months. One of the Dragons has pointed out that every amazing house they’ve lived in has been distilled out of a larger, less-awesome house; it’s simply too hard to predict in advance which subsets of people will gel with one another and with a common spirit or mission. Right now, I feel somewhat fatigued, and may not spearhead a Phase II; if I’d gone into the six months knowing that Phase II would be where the magic happens, I would have spent resources differently.

2. Speaking of spending resources, have a Sabbath. In my role at CFAR, I bear a greater-than-average burden of emotional labor, both internally and at workshops. I may not do a very good job of it, but it is part of my job to be something of an island of stability — to absorb into myself people’s problems, to take blame, to reassure, to cover bases and fill cracks and catch dropped balls. Add into that the explicit role of Commander at Dragon Army, and I’ve essentially been “on call” for emotional labor at all times for the past six months. While setting personal boundaries is a skill I could improve in its own right, having a “nobody in Dragon Army is responsible for anything Dragon Army” day each week would’ve been net positive for everyone.

3. Start those Dragon Army responsibilities early. We held a weekend experiment with ~20 people, and then cohered a group of ~10 out of those. Once that group was finalized, there was something like a six-week period in which we were conducting a housing search, and then also a sort of fragmented three-week move-in period interrupted by some workshops and so forth. A better plan would’ve been to start weekly responsibilities immediately after the weekend experiment (probably one weekly exercise block and one weekly intellectual block), and to take them extremely seriously from the beginning. There are four characters which I will discuss more down below (the White Knight, the Red Knight, the Black Knight, and the Ghost) who I believe would’ve updated away from participation in Dragon Army during such a period, leading to a smaller but tighter-knit group from the beginning.

4. Build in a better off-ramp. I really, really, really hoped that people would take seriously their commitment to be fully in for six months. Most did. Some tried, but were otherwise constrained by either a lack of kingship or a lack of prophet-nature. My original reluctance to build an off-ramp was sort of like a startup founder deliberately choosing not having an exit plan; I knew that we could absorb anything that was likely to go wrong and survive for six months, but I underweighted the degree to which that would reduce our ability to stride boldly forward anyway. In retrospect, it would’ve been better to sort out a formal arrangement with e.g. Event Horizon (another local group house) such that there was a place for people to go after three or four months if the engagement didn’t seem to be working out. Desire to prevent people from feeling insecure about their housing led me to make decisions that were suboptimal given the goals of the project; if there had been a smooth and low-cost way to either “nudge people away” or “set people free” (depending on the frame) that would’ve been better.

*I’ll note that some of the Chicken Littles were clamoring for an off-ramp, but they seemed to me to want it solely for the emotional health of the individual participants, who they otherwise predicted would be ensnared in a hellscape of abuse. I think they were right about an off-ramp being good, but for the wrong reasons.

5. Get your shit together, Duncan. There’s a large category labeled “obvious” that I didn’t quite have all in line by our start date of September. As a simple but representative example, we didn’t get formal weekly check-ins off the ground until several months in. There are a lot of little places where I really should “know better,” given my experience running groups of people, and I relied too much on the group’s general competence and my own ability to catch falling balls, instead of taking a week to get all of the ducks in a row.

5b. Be much more deliberate about second-to-second culture creation. I am more convinced than ever that little things like hand gestures, formal call-and- response patterns, jargon, little rituals, and explicitly outlined ways-of-interacting are a huge make-or-break for something like DA, and that I hit them lightly enough that we ended up in “break” territory. I recommend to my past self that he set aside a full week to brainstorm and memorize a set of cultural tools and norms akin to what a person learns when they first step into a serious martial arts academy, and to relentlessly and unashamedly hammer them in from Day 1. First people feel weird, then they feel comfortable, then they feel proud; there’s a lot of low-hanging fruit here and we ate almost none of it (although even tiny little bits of it in the final month have proven delicious).

6. Relatedly, give more orders. I didn’t give enough orders. I didn’t provide as much structure as half the house showed up hoping for, in part to avoid transgressing with the other half of the house. But assuming that all of this advice gets transmitted at the same time, that other half would have a way to tap out, and the rest would get the Actual Army Experience They Wanted. Instead, there were lots of places where I failed to pass the double illusion of transparency, because we didn’t have enough practice distinguishing orders from suggestions from ideas from discussion. Past Duncan should start off giving lots of orders about lots of things, and iterate by pruning the tree via feedback, as opposed to being light-touch and then regretting his lack of a strong norm of command that he could lean on.

(There’s a thing in here that’s between 4, 5b, and 6, which is having more explicit atonement- and loop-closing-norms. We had house norms of doing pushups and donating $5, to close the loop on transgressions both objective and perceived, and this was nowhere near a big enough stick. There were times when it seemed like the only larger option was expulsion, and this was Dumb Of Me to have not solved before day 1.)

7. Fewer things better done. In the first ~six weeks, we got up together three times a week and exercised. It was awesome, and we had 100% reliability minus actual illness or out-of-town-ness; everybody gave it their all and showed up. It was costly, though, in that the coordination challenge was difficult (e.g. one Dragon has quality- of-life issues from getting up early, and another from staying up late, and another from a long commute), and that it didn’t quite meet the individual fitness goals of ~half the house. So we declared “victory” and moved on to other experiments.

In retrospect, we shouldn’t’ve. House exercise + our regular house dinner were cornerstones, foundations, and trading some of that time back in for yet more scattered and fragmented experiments was a mistake. Instead, we should’ve started with something like four reliable, followed-through commitments (probably something like an exercise slot, a family/togetherness slot, a study/intellectual progress slot, and either a second study slot or some kind of performance or get-stuff-done slot), and kept them solid for the full six months. This would’ve significantly reduced our ability to run a lot of the other experiments we ran, but ultimately I think it would’ve produced a) more satisfaction (once we got past the unpleasant valley and sank into making-the-most-of-the-frame) and b) more actual progress (since instead we sort of kept changing the target after two or three steps and in at least some ways never really “went anywhere”). An example of this is a meta-experiment in running three-week-long experiments with clear check-ins and debriefs, which ran for a total of three overlapping experiments and then got dropped.

8. Actually check in on the ~20h/wk requirement. One Dragon, late in the game, asked outright “So, do we actually have standards, or not?” and the answer was “No, and it’s my fault.” By trying to hold a meta standard of doing-all-the-things-to-a- certain-level-of-good, I multiplied the difficulty of common knowledge and coordination and self-evaluation on all sorts of scattered things going on. The simplest starting point would’ve been “are you actually spending twenty hours per week on stuff that can defensibly be described as house stuff?” which would’ve led into “how’s the stuff that you’re doing going, specifically?” I think that would’ve very early on revealed some persistent differences in how Dragons were conceiving of their relationship to the house, to their goals, and to one another, and provided handholds and levers for me to help or funnel resources their way (or gently guide them out of the experiment).

9. Speaking of persistent differences …

9a. Do something about the White Knight. The White Knight is a role rather than a specific person; a role that various Dragons stepped into at various times (although there was one Dragon who was the White Knight something like an order of magnitude more often than any other). The White Knight is a character who has looked at what’s going on, built a model, decided that they understand the Rules, and has begun to take confident action in accordance with those Rules. In particular, the White Knight has decided that the time to choose stag is obvious, and is already common knowledge/has the Schelling nature.

The White Knight is often wrong. Furthermore, other people often don’t notice that the White Knight is assuming that everyone knows it’s time to choose stag. So the White Knight burns resources over and over again, and feels defected on again and again, and meanwhile the other Dragons feel judged and found wanting according to a standard they never agreed to (remember, choosing rabbit should be the Schelling option, according to me), and the whole thing is very rough for everyone.

I saw this unfolding, took steps to point it out and to ameliorate it, but I didn’t do enough by half. The main thing that the White Knight needs to hear from the king (Commander) is “No. This is not the time or place to choose stag, whatever your preconceptions. If you choose stag now, you will lose resources, and you will feel betrayed, and some part of you will hold your fellow Dragons accountable for it and that’s not fair to them, either.”

I suspect that most Dragons playing the role of the White Knight could’ve heard that message. I suspect that any who couldn’t, or who kept making the mistake, should’ve either had that specific thing turned into one of their growth targets, or else distanced themselves from the experiment. Unfortunately, I didn’t have the model to get it done until about a month ago.

9b. Do something about the Red Knight. Like the White Knight the Red Knight is a role or character more than a real person. The Red Knight is a proponent of uncertainty, with an acute awareness of black swans. The Red Knight is a defender of individual freedom and personal sovereignty, and is wary of making commitments that might prove constraining or coercive. e Red Knight lives in fear of the White Knight’s judgment, or more specifically, in fear of being exposed to and vulnerable to the White Knight’s judgment. It’s not that the Red Knight is afraid of the White Knight so much as that the Red Knight wants to avoid the argument. The Red Knight does not want to be misunderstood about what it’s agreed to or how far it’s willing to go, and so it defaults to “no” over “yes.”

On the very first evening of Dragon Army, I attempted to install and get ratification on a pair of commitments, that I would then consider the house to have given me permission to hold them accountable to. The idea was that these commitments would be levers that the Dragons would allow me to install in them, such that if I tugged on them (“hey, I think you’re failing to live up to commitment one, here”), they would treat that tug as a reminder of principles they wanted to be held accountable for, rather than an outside attempt to coerce and constrain and compel.

These commitments were easily in line with the Charter and the principles we’d lived in our weekend experiment. However, the Red Knight objected, citing that they were too vague and open-ended. My specific advice to past Duncan is, go ahead and draw a clear line and insist on installing the principles, and either reassure the Red Knight in private, or nudge the Red Knight out. My more general advice is “err on the side of upsetting the Red Knight rather than on the side of cohering no structure at all and being afraid to command for the sake of the Red Knight’s sense of freedom.”

The Red Knight doesn’t belong in Dragon Army; if it shows up, it does so in full awareness that local laws apply. (In addition to sacrificing my own goals, by bending over backwards to accommodate the Red Knight, I was also doing the Red Knight itself a disservice — something like infantilizing it or condescending to it, implying that it wasn’t capable of drawing its own boundaries and honoring its own self.)

9c. Do something about the Black Knight. The Black Knight is a mercenary; it has goals of its own, and it’s only allied with you because your goals align with them. This does not mean that the Black Knight is dishonorable; the Black Knight is perfectly aware of the instrumental value of its reputation, and of the power of mutually beneficial coordination, and it is every bit as honest as the White or Red Knights.

Often, the White Knight (once burnt out) becomes the Black Knight. Where the White Knight pours out its heart and soul, the Black Knight husbands its resources. Its cooperation with you is transactional, and measured, and tenuous; it is a Rules Lawyer and a min-maxer, a satisficer rather than an optimizer. It is not With You In Spirit; it is not willing to set aside its sovereignty or its judgment; it is not urge-propagated top-to-bottom on the value of commitment and surrender. It will follow through on its obligations, and will dutifully bow to the orders of the king, but it is not enthusiastic. It seeks convincing argument at exactly the moment that what is required is easygoing optimism; it maintains professional distance at exactly the moment that what is required is a breaking down of walls into bridges.

The Black Knight also doesn’t belong in Dragon Army — or rather, it doesn’t belong in the final Dragon Army. In the world where the first six months are an extended trial period, the Black Knight is an incredibly valuable check upon the excesses of optimism — distilling and refining and iterating and improving, questioning what is needed, what is worthwhile, what is good.

But in order to do the true, steel, Platonic ideal of Dragon Army or something like it, the Black Knight needs to set aside its armor. The Black Knight is the most likely to erroneously choose rabbit, just as the White Knight is the most likely to erroneously choose stag. It’s hard to feel secure in your relation to a Black Knight; everything feels like it’s being weighed and judged (because it is); every action feels like it needs to be justified; it feels like you never have credit in the bank (or rather, like you have credit in the bank, but you’re close to the limit always and you’re always having to budget and you can never relax). That just … doesn’t work, if you’re trying to build a tight-knit team.

9d. Do something about the Ghost. The Ghost is just what it sounds like — the character that isn’t there. Sometimes it’s shut away in its room with the door closed. Sometimes it’s o at work for fourteen hours a day. Sometimes it’s absent from the house for days at a time without anyone else ever really knowing why.

I don’t know if there’s any special advice I’d give myself about the Ghost separate from what’s already been said — if there are fewer and clearer expectations better followed up on by me, and if there’s a pre-move-in period where people can find out if they t with those expectations better, and if there’s a gentle o-ramp, then I suspect the Ghost is less of a problem. e main thing I’d say, I suppose, is to draw a clear line (even if it’s arbitrary). Many of the Dragons who stepped into the role of the Ghost for a time did so softly and gradually, and it never felt like this level of absence was Notably Different from the previous level, in a paradox-of-the-heap sort of way. Set a bar, and set a gradient around that bar, and stay in contact.

10. Put more data on the walls. Shifting gears a little — I’d tell past Duncan to start some kind of permanent record-keeping. We kept a LOT of data out on the walls, in public spaces, but it was always this experiment or that experiment — scattershot and inconsistent. If there’d been one wall that we’d slowly covered with sharpie, or one dry erase board that had all ~180 days of the experiment drawn on it, I predict this would’ve had powerfully focusing effects.

11. Do more menial work. In particular, regular housekeeping. Some time around the halfway mark, we started doing “effort exchange,” which is loosely based on a much-less-politically-correct thing you can ask Andrew Critch about. During EE, pairs of Dragons would spend ~30min being minions for a third Dragon, who would then serve as minion for each of them in turn. So, for about half an hour or so, you would have access to 3x your usual brainpower, 3x your usual amount of experience and perspective, 3x your usual number of hands.

EE was awesome; it is by far the highest contender for continued use in future versions of Dragon Army, even those potential versions which lower the bar and become something like a decentralized Dragon Corps. I’d like to see a version of it that’s more ambitious, and has people learning to use one another in long-term and intellectually complex ways, but the evolved norm of Dragons-pitching-in-on- housework-and-chores is good in its own way. Past Duncan should set something like a 30min housekeeping time thrice per week, where we’re all pitching in together on a very objectively measurable task that’s a little bit unpleasant but ultimately satisfying.

12. Do something consistent on Tuesday evenings. Tuesdays were our “family dinner” … each week, one or two or three Dragons would buy food with the house budget and cook for everyone, and then we’d sit down and eat together, and then we’d do …

… something. Often there was Circling, often there was a sort of logistical check-in, often there were one-on-one conversations, often there was some sort of middle-school-homeroom-esque activity.

The variety was good, but I’d have something be consistent week in and week out. Like, perhaps we eat dinner and then there’s always the same 30-minute reflection on the week, in the same format, before we move on to Commander’s Choice. is is an example of how, in general, there were too few reliable, predictable guardrails.

13. Make Dragons check in with each other. This was less of a concern back when we were doing exercise thrice a week, but once we stopped, I should’ve instituted some system by which each edge of the graph received regular nourishment.

14. Have the house fund be 25% instead of 10%. We asked too little of ourselves, and were constantly dipping into particular Dragons’ generosity rather than just having a house fund that was sufficient and abundant.

15. Have more events that pulled in outsiders. I don’t think we at all ran into the failure mode of being too insular and cultish; if anything, we didn’t pull together tightly enough, and many people (especially in the last two months) ended up Ghosting back toward their prior relationships almost entirely.

So this is not about being more transparent to outsiders as a check against cultural abuse, or abuse from me. Rather, it’s like how CFAR, in its early days, experienced huge growth from forcing itself to run test sessions every single Saturday, with real actual humans showing up and expecting to receive content. If we’d put on more of a show — if we’d precommitted to putting on more of a show — I suspect we’d have more quickly and seriously zeroed in on “okay, are we actually achieving what we’re here to achieve, though?” Sort of like measuring yourself every day, if you’re trying to change weight — the esteem of other rationalists in our neighborhood is actually a fairly tight proxy for the kind of impressive we were trying to be, and if we’d exposed ourselves to it more, we would’ve progressed more and faster.

16. Build in more fights, more elbow-bumping, more critique and criticism, more of a culture of hormesis. Hormesis refers to the fact that there’s a sweet spot of stress where things flourish — too little, and they atrophy; too much, and they wither. I think Dragon Army was insufficiently blunt with itself; insufficiently daring-to-expect-greatness-from-one-another. I take the primary blame for this, as the setter of culture. In particular, there was a point four months in where I wished there were people being openly disappointed in me, but I didn’t have the button to press which said “Hey — how about you judge me as if it actually mattered to you how good I am?”

17. Everything else. Let’s just go ahead and assume that there’ll be things I’ve forgotten, and will want to add in the comments below; let’s just go ahead and assume that some of the questions which you all bring to the table will spark other thoughts in me about how I wish I’d started off differently. My current estimate is that the above seventeen points represent something like 90% of the missing magic (my S1 said it was 95% and I adjusted downward) —

(Oh. Make more bets.)

— and the remaining 10% may or may not be worth going into in detail. Note that I’m going to follow the thread of aliveness re: requests for more information on what specifically went on in the house; in some cases, I’ll feel like responding in depth and with specifics, and in other cases I just won’t. There isn’t anything about the house that’s secret, so if you don’t get what you’re looking for from me, feel free to try pinging other Dragons for their takes; I don’t think there’s anyone in the house that would be sufficiently uninformed or uncharitable that I’d disagree with the data they’d give you.

Happy hunting,
 — Duncan